Smart-Emu5581

joined 10 months ago
sorted by: new top controversial old
[D] how to explain why RL is difficult to someone who knows nothing about it? in c/machinelearning@academy.garden
[–] Smart-Emu5581@alien.top 1 points 9 months ago

Most problems in machine learning have learning of the form "you did something wrong, here is what you should have done instead". RL has learning of the form "you did something wrong, but I'm not telling you what you should have done instead". Imagine trying to learn anything difficult like that.

[D] how to explain why RL is difficult to someone who knows nothing about it? in c/machinelearning@academy.garden
[–] Smart-Emu5581@alien.top 1 points 9 months ago

"I'm millions of times faster than god was" is such a fantastic take. I'm going to remember that one.

[P] Comgra: A library for debugging and understanding neural networks in c/machinelearning@academy.garden
[–] Smart-Emu5581@alien.top 1 points 10 months ago

Yes. The overhead used depends on how often you make it store the data and in how much detail. Both of this is configurable. I find the overhead to be negligible in practice.

[P] Comgra: A library for debugging and understanding neural networks in c/machinelearning@academy.garden
[P] Comgra: A library for debugging and understanding neural networks in c/machinelearning@academy.garden
[–] Smart-Emu5581@alien.top 1 points 10 months ago (1 children)

Mechanistic Interpretability

It's primarily intended for debugging, but it can also help with mechanistic interpretability. Being able to see the internals of your network for any input and at different stages of training can help a lot with understanding what's going on.

1
[P] Comgra: A library for debugging and understanding neural networks (alien.top)
 

I'm a machine learning engineer and researcher. I got fed up with how difficult it is to understand why neural networks behave the way they do, so i wrote a library to help with it.

Comgra (computation graph analysis) is a library you can use with pytorch to extract all the tensor data you care about and visualize it graphically in a browser.

This allows for a much more detailed analysis of what is happening than the usual approach of using tensorboard. You can go investigate tensors as training proceeds, drill down into individual neurons, inspect single data sets that are of special interest to you, track gradients, compare statistics between different training runs, and more.

This tool has saved me a ton of time in my research by letting me check my hypotheses much more quickly than normal and by helping me understand how the different parts of my network really interact.

I hope this tool can save other people just as much time as it did me. I'm also open for suggestions on how to improve it further: Since I'm already gathering and visualizing a lot of network information, adding more automated analysis would not be much extra work.

1
[P] Comgra: A library for debugging and understanding neural networks (alien.top)
 

The newest version of Comgra is out! Thanks to everyone for all your feedback!

Comgra (computation graph analysis) is a library for pytorch to visualize and explore your computation graph in detail. The GUI allows you to look at your data from many different angles, to view examples, check summary statistics, investigate gradients, trace bugs through the network step-by-step, and more.

Use Tensorboard to get an overview of your model, and use Comgra for a deep-dive.

My goal is to make this the go-to library used both by novices who want to understand what's going on under the hood, and by researchers in neural architecture design who need to know all the details.